Research > reasoning capabilities (https://openai.com/o1/)

I led the 50+ person core technical org at Covariant, a series-C company that has raised over $200m in funding. Covariant is building universal AI that enables robots to see, reason, and act on the world around them. My team comprised all AI research and software engineering across US and China.I also served as the overall technical lead of Covariant’s software during my tenure with the company. I set the strategic direction and designed most of the major advances during this time. I wrote code on a regular basis throughout my tenure. Under my leadership, Covariant’s technology for complete robotics solutions evolved from a prototype to a production-grade scalable platform.During my tenure, my team delivered: • Continued improvements to the Covariant Brain, a set of foundational AI models to power universal robotics capabilities spanning areas such as computer vision and motion planning. The platform was used for both first party and third party robotics applications through partners like KNAPP. • A software platform to rapidly and robustly build robotics applications to solve logistics use cases such as induction, sorting (putwall), and kitting. • Shipped complete robotics solutions for our logistics customers, including leading players such as Radial and Otto, to meet materials handling workflow needs in warehouses across US and Europe.

I was the engineering lead of an applied research team working on new frontiers in recommender systems. • Led a team developing an agent-based simulation platform primarily used for recommender system R&D. • Lead responsible for productionizing several factorized and neural deep-retrieval systems widely used in Google recommender systems. • Developed a platform for algorithms that execute multiple distributed TensorFlow phases with hard synchronization barriers at their boundaries. This platform drives most factorized deep retrieval systems at Google. • Championed, implemented, and launched an API change to TensorFlow that improves its semantic consistency by disallowing integer gradients. • Invented a hard-negative mining system for neural two-tower models, commonly used in recommender candidate retrieval. • Parallelized Metropolis-Hasting Monte Carlo algorithms, used to implement a high-performance Ising model on TPU. • Collaborated on an unsupervised learning system for entity resolution.

I led the development of a bespoke highly-scalable detection infrastructure platform during a period of time when the Security team grew from 30 to 800 people. • Managed a team of five engineers. • Invented and built a system for big-data streaming dataflow and temporal logic that became the core of Google's security & privacy detection. • Built (with others) a temporal graph store (XX B edges, X PB) used to correlate disparate entities during security and privacy investigations. • Co-led a project to use deep neural nets for risk-based authorization, anomaly detection, and ACL reduction.

I worked on the Force.com platform, which allows folks to develop their own custom apps in the salesforce.com cloud. Specifically, I worked on Apex Code, the programming language that allows developers to code the logic for their apps, and VisualForce, the component-based user-interface framework for the platform.

MobiTV provides a platform for content delivery to end users over mobile and broadband networks, including live television, video on demand, premium content, satellite radio, and music, interfacing with hundreds of content providers, broadcast and cable television networks, and major music labels, with over 6 million paid subscribers across multiple carrier networks.I worked in the CTO's office. My job was to understand the strategic, technical, and intellectual property aspects of forward-looking projects, to architect and design technical solutions for them, to develop proofs of concept for these designs, to write and review patent proposals, and ultimately to transfer the projects to engineering for implementation and deployment into production. This work involved all aspects of a large scale heterogeneous homegrown software service, including advertising, content management, media delivery, security, and mobile client software.

We were five people building a social networking application for internet and mobile phones. I participated in product development, platform selection, rapid prototyping, and preparation for a financing pitch.

I helped to conceive, design, and implement a groundbreaking new animation system built from scratch. I was responsible for the computational core of the system, including dependency analysis, automatic incremental computation, caching, and generic iterative solvers. I am the sole inventor on several patents covering this system.I led character optimization for Ratatouille, a feature film. I used numerical analysis and caching to speed up articulated characters. I redesigned a free-form deformation module to be optimized for posed projections and shear correction. I optimized a Powell solver used for non-linear inverse kinematics. I achieved a 20x speedup for interactive animation.I improved the character articulation tool in use since Toy Story 2. I invented a patented caching algorithm optimized for interactive animation within a frame. I enhanced the referencing system used for authoring characters.

I led my eight direct reports to develop of an SQL-based distributed network monitoring tool. I was the release manager for various products. I worked closely with product management, QA, technical publications, marketing, sales, and technical support.

I designed a typed structured remote query language for a level-4 router for live media streams. I designed a generic XML XPath-based transformation language as the engine of a flexible and extensible Schema-based validating network-configuration engine. I was the lead developer of several network monitoring tools.

I developed a certifying x86 native-code Java compiler along with an automatic theorem prover. The theorem prover runs on remote clients to ensure the Java safety properties of the certified x86 native code output by the compiler.

I won a National Science Foundation CAREER Award (see below). I co-designed a Java API called Triveni for the compositional construction of concurrent event-based reactive systems.

I instigated technology transfer of Ph.D. research into the Verisoft software verification tool. I designed and implemented a symbolic execution/analysis tool for C programs, similar to Microsoft’s PREfix tool.

I invented some of the first non-trivial abstract interpretations for the semantics-based automatic analysis of concurrent programs.

I invented some of the first non-trivial abstract interpretations for the semantics-based automatic analysis of concurrent programs.

I researched semantics-based static analysis of concurrent programming languages.

I invented the first known efficient most-specific-subsumer (MSS) algorithm for natural-language term-based classification.